feat: support flat repos #90
Closed
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
See bazelbuild/bazel#20369 While working on refactoring, I kept hitting rebase conflicts due to issues with the MODULE lock. I guess it's because the Bazel version in e2e tests is much lower than the current one with the fix (7.2) but still, I don't think it adds much to have the lock in e2e testing.
PR GoogleContainerTools#73 added the `_resolve` and this breaks the buildozer fix / autofix
It seemed like 75afff9 in GoogleContainerTools#47 added the new locks but as new files, that is, the old ones were left behind.
Add support for MODULE.bazel to the lock script and avoid printing an unnecessary (and annoying 😅) error when building in a "modern repo".
Add a `nolock` attribute to avoid getting annoying DEBUG messages for repos that we explicitly want to run without a lock.
* make apt/tests more readable by factoring out the parameters * add a "test suite macro" in each test file that group all of the unit tests in the file and prepends a "test suite prefix". IMHO this is better than using `unittest.suite` because we provide better naming than the automated `_test<NUMBER>` plus these better names are actual targets that can be executed one-by-one by name.
Add other nolock tests to exercise the package repos (the templates, etc).
* separate the script into a template file so it's easier to shellcheck
and syntax highlight in editors.
* shellcheck the script and remove all SC2086 warnings
("Double quote to prevent globbing and word splitting")
* improve the buildozer help messages in the copy.sh script:
* reduce duplication of buildozer command
* add a more clear autofix bazel run command that can be easily
copy-pasted
* change some of the variable names in the copy.sh template for longer,
easier to understand names (repo_name >> name, lock_label >> label)
* move repo_name and workspace_relative_path into variables to reduce
line length and improve readability
While working with some flaky mirrors and trying to figure out why they were failing I found the _fetch_package_index code a bit hard to follow so here's my attempt at streamlining it a bit: * Change the general flow of the for-loop so that we can directly set the reasons for failure in failed_attempts * Remove both integrity as an argument and as a return value since neither is ever used. * return the content of the Packages index instead of the path, since we already have the repository_context. This way we don't need rctx anywhere else. * Reword failure messages adding more context and debug information * Shorter lines and templated strings, trying to make the code easier to read and follow.
* remove the `state` "intermediary `struct`" in `package_resolution.bzl` since it wasn't used / needed. * refactor and move _set_dict from `util.bzl` to a `_package_set` method * use `dict .get()` with default values instead of the "nested `if`s" * renamed `package()` to `package_get` and make it return all package versions when the version is not specified so we can remove `_package_versions` * reordeder `(name, version, arch)` args to match the order of the index keys `(arch, name, version)`
Add testing for package_index mocking the external / side effects (downloads, decompression, etc).
The package resolution debugging that e.g. checks the package dependencies should all be within package_resolution.bzl _resolve_all() and not "leak out" returning the information that's only needed for debugging / logging. Also: * reduce the verbosity of the optional dependencies warning by just printing one message per root package instead of one per package. * break up the long lines to build the error messages and remove the "# buildifier: disable=print".
* move version constraint parsing from package_resolution to its own _parse_version_and_constraint method in version.bzl * refactor _version_relop into a compare method in version.bzl plus a VERSION_OPERATORS dict so that (1) we use the operator strings everywhere and (2) we can use the keys to validate the operators.
Previously we had:
```starlark
pkgindex = package_index.new(rctx, sources = sources, archs = manifest["archs"])
pkgresolution = package_resolution.new(index = pkgindex)
```
And none of the code of package_resolution was used anywhere but in
resolve.bzl and after initializing the `pkgindex`.
Also, it makes sense since we are building the index from the manifest
and once we have the index we use it to resolve the packages and
populate the lock.
Cleanup resolve.bzl and package_index.bzl by moving all of the manifest functionality to a separate manifest.bzl file where we now do all of the work to generate the lock: manifest parsing, validation and the package index and resolution. IMHO this is how it should be because the lock is the "frozen state" of the manifest. * _parse() parses the YAML * _from_dict validates the manifest dict and does the rest of the changes that we need to produce a manifest struct * add extra validation for e.g. duplicated architectures * _lock is the only method that's exposed to the outside and it encapsulates all of the other parts, calling _from_dict and all of the package index and resolution, to produce the lock file. * move get_dupes to util.bzl * refactor the "source" struct into the new manifest where we can now centralize a lot of the structure and logic spread across multiple parts of the code. * remove yq_toolchain_prefix since it's always "yq" and, looking at GH code search, this seems to be a copy-paste leftover from rules_js (or the other way around)... the code is always the same and it never receives a string different from "yq".
* move all of the "package logic" to pkg.bzl
* The v2 lockfile format:
* doesn't need the fast_package_lookup dict because it's already using
a dict to store the packages.
* has the dependencies sorted so the lockfile now has stable
serialization and the diffs of the lock are actually usable and
useful to compare with the changes to the manifest.
* removes the package and dependency key from the lockfile, now it's
done via an external function (make_deb_import_key in
deb_import.bzl)
* Remove add_package_dependency from the lockfile API. Now, the package
dependencies are passed as an argument to add_package. This way, the
lockfile functionality is fully contained in lockfile.bzl and e.g. we
can remove the "consistency checks" that were only needed because
users could forget to add the dependency as a package to the lockfile.
* Ensure backwards-compatibility by internally converting lock v1 to v2.
Also, when a lock is set and it's in v1 format, there's a reminder
that encourages the users to run @repo//:lock to update the lockfile
format.
By separating the migration from the previous commit we get to 1. in the previous commit, run all tests with the new code while locks are still v1 2. update the locks n this commit to V2 so we can then re-run all tests in the final state.
…t.bzl Refactor the package repo templates into their own methods and massively cleanup the `for`-loop in `_deb_package_index_impl`. IMHO overall now there's a much better and clear separation of concerns between the "index repo" (`apt/private/index.bzl`) and the "package repos" (`apt/private/deb_import.bzl`).
jjmaestro
force-pushed
the
refactor-mega-refactor-plus-feat-support-flat-repos
branch
from
d678224
to
c920433
Compare
Fixes issue GoogleContainerTools#56 Follow-up and credit to @alexconrey (PR GoogleContainerTools#55), @ericlchen1 (PR GoogleContainerTools#64) and @benmccown (PR GoogleContainerTools#67) for their work on similar PRs that I've reviewed and drawn some inspiration to create "one 💍 PR to merge them all" 😅 Problem: Debian has two types of repos: "canonical" and "flat". Each has a different sources.list syntax: "canonical": ``` deb uri distribution [component1] [component2] [...] ``` (see https://wiki.debian.org/DebianRepository/Format#Overview) flat: ``` deb uri directory/ ``` (see https://wiki.debian.org/DebianRepository/Format#Flat_Repository_Format) A flat repository does not use the dists hierarchy of directories, and instead places meta index and indices directly into the archive root (or some part below it) Thus, the URL logic in _fetch_package_index() is incorrect for these repos and it always fails to fetch the Package index. Solution: Just use the Debian sources.list convention in the 'sources' section of the manifest to add canonical and flat repos. Depending on whether the channel has one directory that ends in '/' or a (dist, component, ...) structure the _fetch_package_index and other internal logic will know whether the source is a canonical or a flat repo. For example: ``` version: 1 sources: # canonical repo - channel: bullseye main contrib url: https://snapshot-cloudflare.debian.org/archive/debian/20240210T223313Z # flat repos, note the trailing '/' and the lack of distribution or components - channel: bullseye-cran40/ url: https://cloud.r-project.org/bin/linux/debian - channel: ubuntu2404/x86_64/ url: https://developer.download.nvidia.com/compute/cuda/repos archs: - amd64 packages: - bash - r-mathlib - nvidia-container-toolkit-base ``` Disregarding the "mixing" of Ubuntu and Debian repos for the purpose of the example, this manifest shows that you can mix canonical and flat repos and you can mix multiarch and single-arch flat repos and canonical repos. You will still have the same problems as before with packages that only exist for one architecture and/or repos that only support one architecture. In those cases, simply separate the repos and packages into their own manifests. NOTE: The NVIDIA CUDA repos don't follow Debian specs and have issues with the package filenames. This is addressed in a separate commit.
Although the Debian repo spec for 'Filename' (see https://wiki.debian.org/DebianRepository/Format#Filename) clearly says that 'Filename' should be relative to the base directory of the repo and should be in canonical form (i.e. without '.' or '..') there are cases where this is not honored. In those cases we try to work around this by assuming 'Filename' is relative to the sources.list directory/ so we combine them and normalize the new 'Filename' path. Note that, so far, only the NVIDIA CUDA repos needed this workaround so maybe this heuristic will break for other repos that don't conform to the Debian repo spec.
jjmaestro
force-pushed
the
refactor-mega-refactor-plus-feat-support-flat-repos
branch
from
c920433
to
bbf5d62
Compare
Closed
jjmaestro
deleted the
refactor-mega-refactor-plus-feat-support-flat-repos
branch
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Note
Stacked on top of #89
Fixes issue #56
Follow-up and credit to @alexconrey (PR #55), @ericlchen1 (PR #64) and @benmccown (PR #67) for their work on similar PRs that I've reviewed and drawn some inspiration to create "one 💍 PR to merge them all" 😅
Problem
Debian has two types of repos: "canonical" and "flat". Each has a different sources.list syntax:
"canonical": (see https://wiki.debian.org/DebianRepository/Format#Overview)
flat: (see https://wiki.debian.org/DebianRepository/Format#Flat_Repository_Format)
Per the spec,
Thus, the URL logic in
_fetch_package_index()is incorrect for these repos and it always fails to fetch the Package index.Solution
Just use the Debian sources.list convention in the
sourcessection of the manifest to add canonical and flat repos. Depending on whether the channel has one directory that ends in'/'or a(dist, component, ...)structure the_fetch_package_index ()and other internal logic will know whether the source is a canonical or a flat repo.For example:
Disregarding the "mixing" of Ubuntu and Debian repos for the purpose of the example, this manifest shows that you can mix canonical and flat repos and you can mix multiarch and single-arch flat repos and canonical repos.
You will still have the same problems as before with packages that only exist for one architecture and/or repos that only support one architecture. In those cases, simply separate the repos and packages into their own manifests.
Note
This PR also fixes an issue with NVIDIA CUDA flat repos that don't follow the Debian repo spec and have invalid
'Filename'paths.The Debian repo spec for
'Filename'says:However, there are cases where this is not honored. In those cases we try to work around this by assuming
'Filename'is relative to thesources.listdirectory/so we combine them and normalize the new'Filename'path.Note that, so far, only the NVIDIA CUDA repos needed this workaround so maybe this heuristic will break for other repos that don't conform to the Debian repo spec.